Explainable Recommendation Algorithms Based on Historical Reviews Feedback Modelling

可解释推荐系统在提供个性化推荐的同时揭示推荐逻辑，已成为推荐系统领域的重要研究方向。尽管基于Transformer的文本生成技术能够产生流畅的自然语言解释，但其性能高度依赖于数据集的密度，在现实场景中普遍存在的稀疏数据条件下表现受限。针对这一挑战，本文提出一种融合历史评论反馈建模的可解释推荐框架。该框架采用编码器–解码器架构，在特征提取阶段，对用户历史评论和物品关联评论进行深度语义建模，通过注意力机制提取评论中的隐含偏好特征与属性特征。在解释生成阶段，使用多任务联合学习范式，将评分预测任务与文本生成任务进行联合优化。通过在Amazon、Yelp等公开数据集上的实验验证，本模型在评分预测和解释生成任务中均达到了最先进水平，在多数指标上取得了显著的提升。本研究为平衡推荐系统可解释性与数据稀疏性问题提供了新的解决方案。Explainable recommendation systems, which deliver personalized suggestions while revealing their rationale, have become pivotal in recommendation research. Although Transformer-based text generation techniques can produce fluent natural language explanations, their performance heavily relies on dataset density and becomes constrained under sparse data conditions prevalent in real-world scenarios. To address this challenge, this paper proposes a novel explainable recommendation algorithm via historical reviews feedback modeling. The framework adopts an encoder-decoder architecture: during the feature extraction phase, it performs deep semantic modeling of user historical reviews and item-related reviews, extracting implicit preference features and attribute characteristics from reviews through attention mechanisms. In the explanation generation phase, a multi-task joint learning paradigm is employed to jointly optimize the rating prediction task and text generation task. Experimental validation on public datasets including Amazon and Yelp demonstrates that our model achieves state-of-the-art performance in both rating prediction and explanation generation tasks, showing significant improvements across most evaluation metrics. This research provides a new solution for balancing recommendation system explainability with data sparsity challenges.